Radial basis function neural network-based nonparametric estimation approach for missing data reconstruction of non-stationary series

Abstract

In real world, due to various reasons, the data we can acquire is usually incomplete, i.e., a significant number of data can be often missing in a non-stationary time series. Traditional interpolation or estimation methods (e.g., cubic spline) are becoming invalid when the observation interval of the missing data is not small. In this paper we introduced a novel method where a radial basis function (RBF) neural network was particularly designed as an optimal estimator for reconstruction of the missing data, in which several important features of the raw data were chosen as input pattern, and one primary feature was used as the desired output response of the RBF network so as to make it learn enough of the data distribution structure. The experimental simulations on Zooplankton data showed that this method had better performance than other methods such as backpropagation (BP)-based neural network and cubic spline interpolation in the meaning of mean square error and confidence intervals.